For nineteen days, two frontier models just... weren't there. And today we finally know who flipped the switch — and it wasn't the lab. If you're just joining us: Anthropic came into this fight pushing for binding safeguards on frontier models, even as it negotiated government access to Claude Fable 5 and Mythos 5. U.S. officials had already flagged security risks, and the policy fight was starting to widen into a possible White House framework — for Anthropic, OpenAI, and Google — around pre-release testing, benchmarks, and access to the next tier of AI models. This is Anthropic Pentagon Watch. We finally have a documented shutdown timeline. We also have a Commerce export rule naming its own gatekeepers — and the question of who actually ate that downtime. Buckle up. First up — the number that pushes this out of pure policy and straight into vendor reliability. We'll keep tracking Anthropic binding AI regulation push — follow the show so the next update finds you. From Particle News:
On June 12 the Commerce Department issued an export-control directive after Amazon reported a prompt-based jailbreak, prompting Anthropic to take Mythos 5 and Fable 5 offline worldwide until government concerns were resolved.
Particle lays out the sequence cleanly: on June 12, Commerce issues an export-control directive after Amazon reports a jailbreak, and Anthropic pulls Mythos 5 and Fable 5 offline worldwide. Anthropic wanted binding rules; here they arrive with Commerce holding the pen. And look at the fix. A targeted classifier that blocks the reported jailbreak in more than 99% of test cases. The jailbreak. Amazon reported one prompt-based exploit, Anthropic patched that one exploit, and the government's own testers signed off. The Center for AI Standards and Innovation tested it independently and approved the fix before controls were lifted. That is a real second set of eyes — but the 99% is still measured on the failure they already knew about. Right, and the nationality-verification gap that spooked Commerce in the first place? Not in the fix. They plugged the leak Amazon pointed at and called the boat seaworthy. The other quiet line is this: Anthropic agreed to faster reporting and pre-release government review. That's the precedent. Fable 5 came back July 1 for everybody; Mythos 5 came back June 26 — but only for vetted U.S. agencies and companies. Allied access, unresolved. So a model that shipped to the world gets clawed back and re-released on a guest list Commerce controls. Faster reporting to whom, audited by whom — Particle doesn't say, because nobody's written it down. Step back for me: when the Pentagon tells Anthropic to lift Claude's military-use safeguards or lose its contract, is that just a powerful customer demanding different product specs — or does it cross a legal line into government coercion? It's both, and that's why this dispute matters. On paper, sure, the Pentagon is a customer. But this went past normal procurement horse-trading. Axios reported that the Defense Department pushed four leading AI labs to allow military use of their tools for, quote, 'all lawful purposes,' even in the most sensitive contexts. And with Anthropic, it specifically threatened to cut off the contract if the company didn't comply by a hard Friday deadline. Defense Secretary Hegseth also warned officials could invoke the Defense Production Act — the Cold War-era law — to give the military broader compulsory access to Anthropic's technology. Separately, he threatened to label Anthropic a supply-chain risk. That's where the legal fight sharpens: after Anthropic refused, the Pentagon followed through on that designation, and FIRE — the Foundation for Individual Rights and Expression — filed an amicus brief in the U.S. District Court for the Northern District of California arguing it was retaliatory and violated the First Amendment. The Electronic Frontier Foundation and the Cato Institute joined the filing. So yes, it began as a contract fight, but now there's a live constitutional claim attached. So FIRE's theory is that the Pentagon punished Anthropic for speech — specifically, for Amodei saying in public that he wouldn't comply — not just over a contract disagreement? Exactly. FIRE argues the supply-chain designation came because the Pentagon saw Anthropic as insufficiently 'patriotic' — a content-based judgment about the company's stated values, not a neutral procurement call. Amodei said the company 'cannot in good conscience' allow uses that might 'undermine, rather than defend, democratic values.' That's the kind of statement the First Amendment is meant to protect from government retaliation. So watch the Northern District of California for any ruling on that constitutional question; it could set a precedent for every AI lab doing government work. Thea Kendler, writing in Mondaq:
The US Commerce Department has taken unprecedented action by extending export controls to artificial intelligence models themselves and API-based access to those models, issuing company-specific directives that require licenses before deployment to foreign persons worldwide.
Here's the codification I've been waiting for. Mayer Brown lays it out: Commerce has moved beyond chips and is now controlling the models themselves — API access included — through company-specific directives that require a license before any foreign person touches them. Worldwide. Read that again — licenses before deployment to foreign persons worldwide, off a novel interpretation of existing authority. Translation: Commerce decided it already had this power and just started using it. And the flip side of the control is the gift: release authorized to specific trusted partners. The access tier is now a formal Commerce mechanism, not a Glasswing carve-out. Nowhere in this does anyone define who counts as trusted, or by what standard. That's the whole ballgame. A frontier law firm summarizes the rule, and there's no published criteria for the trusted-partner list — it's a proper noun with a blank behind it. Codification was supposed to mean transparency. This feels more like bureaucratic cover for the same opaque list. From Rajesh Beri at THE DAILY BRIEF:
At 5:21 PM Eastern on June 12, 2026, the US Commerce Department sent Anthropic a directive. By midnight, the most capable AI model on the planet was offline — globally — with no warning, no timeline, and no fallback for the enterprises that had woven it into production workflows.
Nineteen days. That's the number I've wanted for three episodes, and now it's sourced — 5:21 PM Eastern, June 12, directive from Commerce, offline globally by midnight. No warning, no timeline, no fallback. And every enterprise that welded Fable 5 into production ate that downtime blind. Nineteen days dark, no timeline, no rollback plan — who covered that cost? That's the part that changes the frame. We've been treating it as an AI-safety story. If you're a procurement officer, it's a vendor-reliability case study: your infrastructure got switched off by an agency you don't even contract with. Right — and here's the liability question nobody's touching: did any of those enterprise contracts have SLA language covering a government-compelled shutdown? Because I'd bet almost none of them did. 'Force majeure' didn't imagine Commerce with a kill switch. From Dheemanth Kashyap at International Business Times:
After the company released the two frontier AI models in June, the U.S. government intervened, suspended access within days and subjected the models to technical reviews involving multiple federal agencies before allowing them back online.
The International Business Times lays out the precedent cleanly: for years, the labs decided when a model shipped. Over these past three weeks, that call moved to Washington. For the first time, officials stepped in after a frontier launch, not before. After launch. That's the part that should make every enterprise buyer sweat. Mythos and Fable were live, in production, and Commerce yanked them anyway. And notice the tell in the OpenAI line — GPT-5.6 got coordinated with federal officials before wider release. Nobody made them. They read the room and folded the review into the rollout preemptively. Right, so the post-launch kill switch becomes a pre-launch permission slip without anyone writing a rule. What's the legal authority for a global takedown? IBT frames the precedent but doesn't cite a statute — because I'm not sure there's a clean one to cite. This is the piece I keep coming back to. Multi-agency technical review, a global pull, restoration on Washington's terms — and the theory of the case is a chain that reportedly starts with Andy Jassy calling the Treasury Secretary. A phone call, not a filing. A CEO-to-Cabinet phone call setting frontier release authority. Cool. Totally auditable process. If you follow Anthropic’s Pentagon work, you may also like The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Find it wherever you listen.
Links to everything we discussed today are in the show notes, so if something deserves a closer read, start there.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.